Fixes the problem with bootstrap trajectories not training for multiple epochs.
The cause of the problem was that callbacks include trajectory state, so they must be reinitialized for each training trajectory. As a result, SKlearnWrapper now expects argument callback_constructors (which is a list of lambda functions that each return a Callback object) instead of callbacks (which was a list of Callback objects).
Fixes the problem with bootstrap trajectories not training for multiple epochs.
The cause of the problem was that callbacks include trajectory state, so they must be reinitialized for each training trajectory. As a result,
SKlearnWrapper
now expects argumentcallback_constructors
(which is a list of lambda functions that each return aCallback
object) instead ofcallbacks
(which was a list ofCallback
objects).